[Numpy-discussion] dot and MKL memory

Mads Ipsen mads.ipsen at gmail.com
Fri Feb 27 18:47:59 EST 2015



On 27/02/15 17:50, Charles R Harris wrote:
>
>
> On Fri, Feb 27, 2015 at 9:02 AM, Mads Ipsen <mads.ipsen at gmail.com
> <mailto:mads.ipsen at gmail.com>> wrote:
>
>     Hi,
>
>     If I build Python 2.7.2 and numpy-1.9.1 and run the following script
>
>     import numpy
>     c = numpy.ones((500000, 4))
>     mat = numpy.identity(4)
>     r = numpy.dot(c, mat)
>
>     the evaluation of the 'dot' increases the memory by app. 35 MB.
>
>     If, in addition, I build numpy-1.9.1 with MKL support, and run the
>     script, the evaluation of the 'dot' increases the memory by app. 450 MB.
>
>     Is the expected?
>
>     Best regards,
>
>     Mads
>
>     specs:
>
>     Ubuntu 12.04
>     ifort (IFORT) 14.0.1 2013100
>     gcc version 4.6.3 (Ubuntu/Linaro 4.6.3-1ubuntu5)
>
>
> No, but I don't know why that is happening with MKL. Can anyone else
> reproduce this?
>
> Chuck
>
>
>
> _______________________________________________
> NumPy-Discussion mailing list
> NumPy-Discussion at scipy.org
> http://mail.scipy.org/mailman/listinfo/numpy-discussion
>

I have now tested and reproduced this on two different Ubuntu 12.04 
boxes. Both were tested with a vanilla Python-2.7.9 install and two 
different MKL setups

*) composer_xe_2015.2.164
    icc (ICC) 15.0.2 20150121
    ifort (IFORT) 15.0.2 20150121

    build: python setup.py build --compiler=intelem
           python setup.py install

*) composer_xe_2013_sp1.1.106
    ifort (IFORT) 14.0.1 20131008
    gcc version 4.6.3

    build: python setup.py config --compiler=unix config_fc 
--fcompiler=intelem install

On both setups I use the following site.cfg

[mkl]
library_dirs = /path_to_intel_composer/mkl/lib/intel64/
include_dirs = /path_to_intel_composer/mkl/lib/intel64/mkl/include/
mkl_libs = mkl_rt
lapack_libs =

If I omit the MKL setup (i.e. build without) the 400 MB memory goes
away.

If I by any means can provide you with some valuable info, please let
me know.

Best regards,

Mads


-- 
+---------------------------------------------------------------------+
| Mads Ipsen                                                          |
+----------------------------------+----------------------------------+
| Overgaden Oven Vandet 106, 4.tv  | phone:              +45-29716388 |
| DK-1415 København K              | email:      mads.ipsen at gmail.com |
| Denmark                          | map  : https://goo.gl/maps/oQ6y6 |
+----------------------------------+----------------------------------+



More information about the NumPy-Discussion mailing list